How Transferable are Herbarium-Field Features in Few-Shot Plant Identification with Triplet Loss?

Due to the limited tropical plant field photos but increasing digitized herbarium specimens, cross-domain plant identification has been employed to investigate the use of herbarium specimens on automated plant identification. However, the domain shift between herbarium and field images makes identifying plant species across these two domains challenging. One of the recent papers shows the superiority of triplet loss in preserving similarities between the crossed domains. Nevertheless, the impact of data deficiency on the performance of this triplet loss model has yet to be studied in depth in this field. Specifically, the transferability of the model features trained from limited cross-domain data to target images in the field domain. Therefore, this paper investigates the robustness of cross-domain plant features learned using this triplet loss metric learning approach compared to the supervised classification approach under general and few-shot experimental settings. Detailed experiments show that the triplet loss metric approach outperformed the supervised classification approach in the few-shot setting and achieved comparable results in the general experimental setting. In addition, the feature dictionary generation schemes composed of various herbarium field feature combinations we proposed boost our models’ performance significantly compared to a single feature type dictionary strategy.